A comparative study and development of a novel deep learning architecture for accelerated identification of microstructure in materials science

Abstract

The properties and processing of materials depend on the microstructure at various stages. Identifying the microstructure is essential for the accelerated development of advanced materials and Material Informatics initiatives. In the current study Convolutional Neural Network (CNN) models such as ResNet, VGG, DenseNet, Inception based, MobileNet based, SVC, Tree based, MLP based are trained and compared for microstructure classification. The models are developed with fewer parameters by decreasing the depth and width of the network. The model based on DenseNet shows high testing, training, and validation accuracy with an optimum number of parameters. The developed model based on DenseNet has a total number of parameters of 421674, including 10,930 non-trainable parameters and 4 Dense blocks, 3 Transition blocks, and the fully connected network. The current model achieved a training accuracy of 97.45%, validation accuracy of 96.26%, and testing accuracy of 94.89%. The developed model is used for classifying the microstructure of high entropy alloys, and results show a good agreement. These state-of-the-art models will accelerate the material development and Material Informatics initiatives.